CVE-2026-34225: open-webui: blind SSRF via image-edit URL fetch

### Summary There is a blind server side request forgery in the functionality that allows editing an image via a prompt. The affected function will perform a GET request on the URL provided by the user. There is no restriction on the domain of the provided URL allowing the local address space to be interacted with. Since the SSRF is blind (the response cannot be read) impact is port scanning of the local network because it can be confirmed if the port is open based on if the GET request failed. ### Details The vulnerability occurs here: https://github.com/open-webui/open-webui/blob/2b26355002064228e9b671339f8f3fb9d1fafa73/backend/open_webui/routers/images.py#L850-L916 Line 911 shows the user provided URL passed to the function `load_url_image`. Within this function on line 883 HTTP/HTTPs URLs are trusted blindly and called asynchronously with `requests.get`. ### PoC The vulnerability can be reproduced with the following curl command: ``` curl -X POST http://localhost:3000/api/v1/images/edit \ -H "Authorization: Bearer <token>" \ -H "Content-Type: application/json" \ -d '{"form_data":{ "image": "<url>", "prompt": "poc"} }' ``` ### Impact Response differentials can be used to port scan the local network: <img width="3016" height="736" alt="image" src="https://github.com/user-attachments/assets/93b4df52-b23c-4ed7-a5fa-9cbedb30091c" /> This can be automated to iterate through the entire port range to determine open ports. If the service running on an open port can be inferred the user may be able to interact with it in a meaningful way if the service offers any state changing GET request endpoints. ### Remediation Restrict provided URLs from local address space.

An attacker who holds any valid, even low-privileged, open-webui account (obtained via credential stuffing, an open self-registration flow, or an intentionally provisioned test user) issues repeated POST requests to /api/v1/images/edit, substituting the image URL with internal targets such as http://10.0.0.5:6379, http://localhost:8500, or http://169.254.169.254/latest/meta-data/. By observing which requests succeed versus fail (or differ in timing), the attacker builds a map of live internal hosts and open ports — Redis, etcd, internal admin panels, cloud metadata — all without ever seeing response bodies. If any discovered service exposes an unauthenticated, state-changing GET endpoint (e.g., a cache-flush or admin-action URL), the attacker can trigger it blind, turning pure reconnaissance into limited unauthorized action against internal infrastructure.